Deep Learning Applied to Ship Image Recognition

View Statistics Email Alert RSS Feed

Information

Details

Project title

Code/計畫編號

MOST109-2221-E019-054

Translated Name/計畫中文名

應用深度學習於船舶影像辨識之研究

Project Coordinator/計畫主持人

Le-Na Chang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Communications, Navigation and Control Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=13531819

Year

2020

Start date/計畫起

01-08-2020

Expected Completion/計畫迄

31-07-2021

Bugetid/研究經費

719千元

ResearchField/研究領域

電信工程

For Taiwan, an island country, maritime transportation and fisheries are essential to the national economy. In recent years, the world has paid more attention to the protection of the marine environment and the sustainable management of fishery resources. However, the increasing demands for international marine transportation promote the development of ships towards larger size, faster speed and automation. Regarding balancing the above two issues and ensuring the safety of Taiwan's waters and coastlines, it is a first priority to manage ships effectively. Thus, in the study, we propose an artificial intelligence approach for “automatic ship image recognition”. The study will be fulfilled in two years. In the first year, “ship detection, ship classification and ship number recognition” will be studied based on the deep learning methods. In the second year, the results of the previous year will be applied to the "ship image recognition system" of Taiwan's intelligent fishing port.Since the performance of deep learning networks is highly related to training data, we will first collect images of various types of ships in Taiwan's waters and port areas with high-sensitivity light and infrared cameras and build a ship image data set for Taiwan. Then, we will explore the effectiveness of the deep learning networks such as YOLOv3 (You only look once v3) and Convolutional Recurrent Neural Network (CRNN) which currently widely used for target recognition, in ship classification and recognition. Here, “TensorFlow” framework will be used to quickly build the deep learning network model. Moreover, to solve the large computational loads required in network training, we will use the high-performance parallel acceleration computing units to improve network efficiency and realize real-time ship recognition. Fishing port management is essential for fishermen's livelihood and safety. At present, the management of Taiwan's fishing ports mostly depends on manpower. In the second year, we will apply the results of the previous year to fishing boat image recognition. The data of fishing boats entering and leaving the port will be accurately recorded. Since the fishing boats moored at the ports will be monitored throughout the day, the efficiency of the fishing port management can be improved. We will first collect fishing boat image of different types and tonnages, and establish a fishing boat image data set. Considering the high density of vessels docked in fishing ports, we will modify the bounding box of the deep learning network in the first year, and reduce the number of network layers and convolution kernels. Finally, these results will be applied to the construction of a fishing boat recognition system, which will provide a low-cost, real-time, accurate intelligent fishing port management. 台灣為海島國家，海運和漁業是國家經濟的命脈。近年來全世界對於海洋環境的保護和漁業資源的永續經營越來越重視，而國際海運需求量卻逐年增加，促使船舶朝向更大型化、快速化及自動化發展。為了兼顧二者並確保我國海域和海岸線安全，對船舶有效監控與管理，實為發展海運和漁業之際刻不容緩的要務。本計畫針對此議題，提出應用人工智慧方法於「船舶影像自動辨識」研究。此計畫將分二年完成，第一年將先結合人工智慧深度學習方法進行「船舶偵測、船舶類型和編號辨識」研究，第二年再將前一年成果應用於「建置台灣智慧漁船影像辨識系統」。因深度學習網路的效能和訓練資料息息相關，我們將先透過高感光倍率攝影機和紅外線攝影機全天候蒐集台灣海域和港區各類型的船舶影像，建立台灣船舶影像數據集。接著，對目前廣泛應用於目標物辨識的深度學習網路，如YOLOv3(You only look once v3)和卷積循環神經網路(Convolutional Recurrent Neural Network，CRNN)，進行模型研究，探討其於船舶特徵萃取、船舶類型和船舶編號辨識的效能。此部分將導入目前被廣為使用的深度學習框架「TensorFlow」以快速建置船舶辨識的深度學習網路模型。並以高效能平行加速運算單元解決網路訓練需大量迭代運算之問題，提高網路效率以及提昇影像偵測速度，達到即時船舶辨識之目的。漁港與漁民的生計和安全息息相關，目前台灣漁港的管理大多依賴人力。第二年研究擬依據前一年船舶影像辨識的成果，更進一步提出「漁船影像辨識」研究，藉由精確紀錄漁船進出港的數據以及對港口停泊漁船之監控，提高漁港的管理效能。研究中將先蒐集不同類型和噸位的漁船影像資料，建立漁船影像數據集。考慮漁港停靠的船隻通常密度較高，因此我們將改良第一年的深度學習網路的邊界框，並進行網路層和卷積核數目優化。最後將此成果應用於建置漁船影像辨識系統，透過24小時的漁船自動辨識，提供兼備低成本、即時、精確的智慧化漁港管理。

Keyword(s)

船舶影像辨識
人工智慧
深度學習
YOLO網路
卷積循環神經網路
智慧化漁港
ship image recognition
artificial intelligence
deep learning
YOLOv3 (You only look once v3)
Convolutional Recurrent Neural Network (CRNN)
intelligent fishing port

DSpace CRIS

Deep Learning Applied to Ship Image Recognition

Details

Description